[IL-William]

Patch 13 of the engine fork that collin pins (dbt-collin decision 0028). A table or CTE read without an alias is now known by its bare name in its own scope.

Until now it registered nothing there, so a qualifier naming it was looked up in the enclosing scopes first:

SELECT reviews.id FROM events AS reviews
WHERE EXISTS (SELECT 1 FROM raw.reviews WHERE reviews.book_id = 1)

put the subquery's filter on events, quoted table or not; with FROM raw.reviews AS reviews it landed right. Since patch 12 (#12) a select-list subquery is analysed too, and collin's tests for correlated subqueries caught it there. Where no alias competed, the qualifier fell through to the bare name taken as a table, without its schema, so SELECT orders.id FROM db.sch.orders read no column at all.

The change

• A TableFactor::Table without an alias, table or CTE, registers its bare name in its scope (Scope::bare_tables), normalized as qualifiers are (normalize_identifier), mapping to its canonical.
• resolve_table_alias and resolve_alias_instance check, innermost scope first: the aliases, the bare names, then the names with no table behind them; the first scope that knows the qualifier answers.
• A bare name two different relations of one scope share resolves to neither and stops the search; the same relation twice keeps it.
• A LATERAL VIEW's name, and the aliases of table functions, PIVOT, UNPIVOT and UNNEST, are registered normalized in the scope (register_subquery_alias), so they shadow an outer alias of their name too. The global alias set keeps them as written, as before.

Left as it is:

• a quoted table's bare key is folded as if unquoted, which matters only for SQL that would not compile on Postgres or Snowflake;
• the qualifier reaches the bare-name lookup unquoted, so a quoted outer alias is folded: FROM events AS "Reviews" WHERE EXISTS (SELECT 1 FROM raw.reviews WHERE "Reviews".kind = 'x') puts the filter on reviews, with a warning that it has no kind, where master placed none; master already fails a quoted alias in the outer select;
• an ordinary alias still matches exactly as written (DuckDB's UNNEST(...) AS U(id) under an outer u), as before this patch.

Tests

• an unaliased table shadows an outer alias of its name, in an EXISTS and in a select-list subquery, quoted and not, on Snowflake and DuckDB;
• an unaliased CTE shadows one too;
• an outer unaliased table is still read through its bare name;
• columns qualified with the bare names of unaliased tables are read;
• a bare name two tables share resolves to neither;
• a LATERAL VIEW name (e and E), a TABLE(GENERATOR(...)) alias and an upper-case UNNEST alias shadow an outer alias.

Each fails without its part of the patch. cargo test -p flowscope-core, cargo fmt --all -- --check and CI's cargo clippy --workspace --locked -- -D warnings pass. The same commit applies unchanged to collin and master.

Measured through collin

With a local [patch] on this branch, collin's main:

Project	Effect
lineage-library	book_ratings (with patch 12), loan_titles and shelf_tags give their lineage as written: nineteen marks go, nothing else moves
collin's tests	the two correlated-subquery tests that failed on patch 12 alone pass; one test that relied on the engine's false alarm for a CTE named like the alias inside it now finds none
Jaffle Shop	cache identical, --indirect too
Shopify	cache identical, --indirect too

🤖 Generated with Claude Code

https://claude.ai/code/session_018LDZwk297SKdJXXLpTFZEh